Performance analysis of remote photoplethysmography deep filtering using long short-term memory neural network

Remote photoplethysmography (rPPG) is a technique developed to estimate heart rate using standard video cameras and ambient light. Due to the multiple sources of noise that deteriorate the quality of the signal, conventional filters such as the bandpass and wavelet-based filters are commonly used. However, after using conventional filters, some alterations remain, but interestingly an experienced eye can easily identify them.

We studied a long short-term memory (LSTM) network in the rPPG filtering task to identify these alterations using many-to-one and many-to-many approaches. We used three public databases in intra-dataset and cross-dataset scenarios, along with different protocols to analyze the performance of the method. We demonstrate how the network can be easily trained with a set of 90 signals totaling around 45 min. On the other hand, we show the stability of the LSTM performance with six state-of-the-art rPPG methods.

This study demonstrates the superiority of the LSTM-based filter experimentally compared with conventional filters in an intra-dataset scenario. For example, we obtain on the VIPL database an MAE of 3.9 bpm, whereas conventional filtering improves performance on the same dataset from 10.3 bpm to 7.7 bpm. The cross-dataset approach presents a dependence in the network related to the average signal-to-noise ratio on the rPPG signals, where the closest signal-to-noise ratio values in the training and testing set the better. Moreover, it was demonstrated that a relatively small amount of data are sufficient to successfully train the network and outperform the results obtained by classical filters. More precisely, we have shown that about 45 min of rPPG signal could be sufficient to train an effective LSTM deep-filter.